Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method for controlling an operation of a power generating unit including at least one of a gas turbine and a steam turbine, the method comprising: operably coupling a control system to the power generating unit for controlling an operation thereof, the control system comprising a disturbance rejection model that includes a network for mapping system inputs to a system output and configured to generate a predicted value for the system output at a future time; training the disturbance rejection model per a training dataset; calculating a confidence metric for the disturbance rejection model; providing, with the confidence metric, a probability that a predicted sign of a value change between the predicted value for the output at the future time and a measured value of the output at a previous time occurring prior to the future time is correct, wherein the predicted sign is a positive sign if the predicted value for the output at the future time increases from the measured value of the output at the previous time; and the predicted sign is a negative sign if the predicted value for the output at the future time decreases from the measured value of the output at the previous time, and controlling the operation of the power generating unit based on the calculated confidence metric.
This invention relates to the field of power generation control, specifically addressing the problem of maintaining stable and efficient operation of power generating units, such as gas turbines and steam turbines, in the presence of disturbances. The method involves operably coupling a control system to the power generating unit. This control system incorporates a disturbance rejection model. The model is designed to map system inputs to a system output and is capable of predicting the system output at a future time. This disturbance rejection model is trained using a dataset. A key aspect of the invention is the calculation of a confidence metric for the trained disturbance rejection model. This confidence metric quantifies the probability that the predicted sign of a change in the system output is correct. Specifically, it assesses the likelihood that the predicted future output will be higher or lower than a measured past output, indicating an increase or decrease, respectively. Finally, the operation of the power generating unit is controlled based on this calculated confidence metric. This allows for adaptive and robust control, ensuring the unit's performance is maintained even when faced with unpredictable operational variations.
2. The method of claim 1 , wherein the training dataset comprising a time series dataset of actual values for the system inputs and the system output as determined from measurements taken of the operation of the power generating unit during a historical operating period, the historical operating period occurring prior to the future time; and wherein the training dataset comprises T samples, the T samples taken at times=[1, 2, . . . ,t, . . . ,T] wherein a time (t), a time (t−1), a time (t−2) denote respective samples in the training dataset where the time (t−2) occurs just prior to the time (t−1) and the time (t−1) occurs just prior to the time (t).
This invention relates to predictive modeling for power generating units, specifically using time-series data to forecast system performance. The method addresses the challenge of accurately predicting future operational states of a power generating unit by leveraging historical operational data. The training dataset consists of time-series measurements taken during a historical operating period before the future time of interest. This dataset includes actual values for system inputs and outputs, collected at discrete time intervals. The samples are sequentially ordered, with each sample (t) following the previous sample (t−1), which in turn follows (t−2), ensuring a chronological representation of system behavior. The dataset contains T samples, indexed from 1 to T, where each sample captures the state of the system at a specific time. This structured approach allows for the development of predictive models that can analyze temporal patterns in the data to forecast future system performance. The method ensures that the training data reflects real-world operational conditions, improving the accuracy and reliability of predictions for power generation systems.
3. The method of claim 2 , wherein the disturbance rejection model comprises a disturbance rejection configuration in which the predicted value made by the disturbance rejection model for the system output at the future time is based upon a predicted value made by the network for the system output at the future time and an error multiplied by a feedback coefficient; and wherein the error comprises a difference between: a predicted value made by the network of the output at the previous time; and the measured value of the system output at the previous time, wherein the measured value is based upon a measurement taken by a sensor disposed in the power generating unit for measuring an operating parameter related to the system output.
This invention relates to a method for improving disturbance rejection in a power generating unit using a predictive model. The method addresses the challenge of accurately predicting system outputs in the presence of disturbances, which can degrade performance and efficiency in power generation systems. The disturbance rejection model is configured to predict future system outputs by combining a network's prediction with an error correction term. The error is calculated as the difference between the network's predicted output at a previous time and the actual measured output, where the measurement is obtained from a sensor monitoring an operating parameter of the power generating unit. The error is then scaled by a feedback coefficient and incorporated into the prediction to enhance accuracy. This approach allows the system to dynamically adjust predictions based on real-time sensor data, improving robustness against disturbances. The method ensures that the predictive model remains accurate even when external factors or system variations introduce deviations, leading to more reliable control and optimization of the power generating unit.
4. The method of claim 3 , wherein the network comprises a neural network that includes multiple layers having nodes, the multiple layers including at least an input layer, an output layer, one or more hidden layers, and forward weight matrixes; and wherein: the input layer comprises a plurality of the nodes, the plurality of the nodes corresponding respectively to the system inputs, wherein each of the plurality of the nodes is configured to receive an input signal relating to a value for a particular one of the system inputs; the output layer comprises at least one of the nodes, the at least one of the nodes corresponding to the system output; the one or more hidden layers are disposed between the input layer and the output layer, each of the one or more hidden layers comprising a plurality of the nodes; and the forward weight matrices comprise connectors that connect the nodes of successive layers of the multiple layers of the neural network and a weight value for each of the connectors; wherein a weight vector defines the weight values for the connectors of the forward weight matrices.
This invention relates to neural network architectures for processing system inputs to generate system outputs. The problem addressed is the efficient and accurate representation of data transformations in neural networks, particularly in systems requiring precise input-output mappings. The neural network includes multiple interconnected layers: an input layer, an output layer, and one or more hidden layers. The input layer consists of nodes, each corresponding to a system input and configured to receive input signals representing specific input values. The output layer contains at least one node corresponding to the system output. Hidden layers, positioned between the input and output layers, each contain multiple nodes. The network uses forward weight matrices to connect nodes across successive layers, with each connection assigned a weight value. These weight values are collectively defined by a weight vector, which governs the strength of connections between nodes. The architecture enables flexible and scalable data processing, allowing for complex input-output relationships to be modeled effectively. The forward weight matrices facilitate the propagation of signals through the network, with the weight vector determining the influence of each connection on the final output. This design is particularly useful in applications requiring high-dimensional data transformations, such as pattern recognition, predictive modeling, and decision-making systems.
5. The method of claim 4 , wherein the confidence metric is based on a probabilistic distribution calculated for the T samples of the training dataset; wherein the system inputs at the time (t−1) and the time (t−2) are described by input vectors, which are denoted, respectively, as an input vector (x t−1 ) and an input vector (x t−2 ); wherein the system output at the time (t−1) is denoted as a system output (d t−1 ); wherein the step of calculating the probabilistic distribution for the predicted value of the system output at the time (t) comprises calculating a mean (μ) and a variance (σ 2 ) that define a normal distribution (N).
This invention relates to predictive modeling systems that evaluate confidence in predicted outputs. The problem addressed is the need for a robust confidence metric in predictive systems, particularly when dealing with time-series data or sequential inputs. The system processes a training dataset containing T samples to generate a probabilistic distribution for predicted values. Input vectors at two prior time steps (t−1 and t−2) are used to predict the system output at the next time step (t). The system output at time (t−1) is denoted as d(t−1). To calculate the confidence metric, the system computes a normal distribution (N) defined by a mean (μ) and variance (σ²) for the predicted value at time (t). This probabilistic approach allows the system to quantify uncertainty in predictions, improving reliability in applications such as forecasting, control systems, or decision-making processes. The method ensures that confidence metrics are derived from statistical properties of the training data, providing a mathematically rigorous basis for evaluating prediction accuracy. The invention enhances predictive systems by incorporating uncertainty estimation, which is critical for applications requiring high reliability.
8. The method of claim 1 , wherein an algorithm for calculating the confidence metric is configured such that the confidence metric comprises: a high probability that the predicted sign of the value change for the system output at the time (t) is correct when the variance of the probabilistic distribution is small compared to a difference between: the predicted value of the system output made by the trained disturbance rejection model at the time (t) and the system output (d t−1 ) at the time (t−1); and a low probability that the predicted sign of the value change for the system output at the time (t) is correct when the variance of the probabilistic distribution is large compared to the difference between: the predicted value of the system output made by the trained disturbance rejection model at the time (t) and the system output (d t−1 ) at the time (t−1).
The invention relates to a method for evaluating the confidence of predictions made by a disturbance rejection model in a control system. The problem addressed is the need to assess the reliability of predicted changes in system output, particularly when dealing with uncertainties in the system's behavior. The method calculates a confidence metric for the predicted sign of a system output change at a given time (t). The confidence metric is derived from a probabilistic distribution associated with the model's prediction. When the variance of this distribution is small relative to the difference between the predicted output at time (t) and the actual output at the previous time (t−1), the confidence metric indicates a high probability that the predicted sign of the change is correct. Conversely, when the variance is large relative to this difference, the confidence metric indicates a low probability that the predicted sign is correct. This approach allows the system to dynamically adjust its confidence in predictions based on the uncertainty in the model's output, improving decision-making in control applications where reliability is critical. The method ensures that predictions are only trusted when the model's uncertainty is sufficiently low compared to the magnitude of the predicted change.
9. The method of claim 5 , wherein, given the probabilistic distribution of the predicted value of the system output at the time (t), the confidence metric comprises a probability calculated via a fraction in which: the numerator of the fraction comprises a sign for: the system output (d t−1 ) at the time (t−1) subtracted from the probabilistic distribution of the system output made by the trained disturbance rejection model at the time (t); and the denominator of the fraction comprises a sign for: the system output (d t−1 ) at the time (t−1) subtracted from the predicted value of the system output made by the trained disturbance rejection model at the time (t) and the system output (d t−1 ) at the time (t−1).
This invention relates to a method for calculating a confidence metric in a disturbance rejection model used for predicting system outputs. The method addresses the challenge of assessing the reliability of predictions in dynamic systems where disturbances can affect output accuracy. The disturbance rejection model is trained to predict system outputs, and the confidence metric quantifies the model's certainty in its predictions by comparing the predicted output with the actual output from the previous time step. The confidence metric is derived from a probabilistic distribution of the predicted system output at a given time (t). It is calculated as a fraction where the numerator represents the difference between the system output at time (t−1) and the probabilistic distribution of the system output predicted by the trained model at time (t). The denominator represents the difference between the system output at time (t−1) and the predicted value of the system output at time (t), normalized by the system output at time (t−1). This approach provides a normalized measure of prediction confidence, accounting for variations in system behavior and disturbances. The method ensures that the confidence metric reflects the model's ability to reject disturbances and accurately predict system outputs.
10. The method of claim 5 , wherein given the probabilistic distribution of the predicted value of the system output at the time (t), which is designated as τ t , the confidence metric comprises a probability (p) that the predicted sign of the value change for the system output is correct; wherein the probability (p) is given by: p ( sign ( τ t - d t - 1 ) sign ( y ( x t - 1 , x t - 2 , d t - 1 , k * , w MAP ) - d t - 1 ) > 0 ) .
This invention relates to probabilistic prediction systems for evaluating the confidence in predicted system output values. The problem addressed is the need to assess the reliability of predicted changes in system outputs, particularly when dealing with probabilistic distributions of predicted values. The invention provides a method to compute a confidence metric that quantifies the probability that the predicted sign of a system output's value change is correct. The method involves analyzing the probabilistic distribution of the predicted system output at a given time (t), denoted as τt. The confidence metric is derived from the probability (p) that the predicted sign of the value change aligns with the actual observed sign. Specifically, the probability (p) is calculated by comparing the sign of the predicted value (τt) with the sign of the difference between the system output (y) at the previous time step (t-1) and a baseline value (dt-1). The system output (y) is determined using a model that incorporates input variables (xt-1, xt-2), the baseline value (dt-1), and learned parameters (k*, wMAP). The method ensures that the confidence metric reflects the likelihood that the predicted direction of change in the system output is accurate, providing a robust measure of prediction reliability.
12. The method of claim 11 , wherein, given the Gaussian Cumulative Distribution (Φ), the probability (p) comprises: p ( sign ( τ t - d t - 1 ) sign ( y ( x t - 1 , x t - 2 , d t - 1 , k * , w MAP ) - d t - 1 ) > 0 ) = 1 2 + 1 2 erf ( y ( x t - 1 , x t - 2 k * , w MAP ) - d t - 1 2 σ 2 ( x t - 1 , x t - 2 , k * , w MAP ) ) where the erƒ( )comprise a Gauss error function.
This invention relates to probabilistic modeling in machine learning, specifically for estimating the likelihood of a binary outcome based on Gaussian distributions. The method addresses the challenge of accurately computing probabilities in systems where outcomes depend on prior states and learned parameters. The core technique involves calculating a probability (p) using a Gaussian Cumulative Distribution Function (Φ) and the error function (erf). The probability is derived from the product of two sign functions and a Gaussian error term. The first sign function compares a threshold (τ) at time t with a decision boundary (d) at time t-1. The second sign function compares a predicted value (y) at time t-1, which depends on prior inputs (x) and learned parameters (k*, w_MAP), with the same decision boundary (d) at time t-1. The probability is then computed as a function of the absolute difference between the predicted value and the decision boundary, scaled by the variance (σ²) of the prediction. This approach enables precise probabilistic reasoning in dynamic systems where decisions are influenced by historical data and learned models. The method is particularly useful in applications requiring real-time decision-making, such as control systems, financial forecasting, or adaptive signal processing.
13. The method of claim 11 , wherein the confidence metric, designated below as η, is computed as: η = 100 ∑ t = 2 T C t , t Φ ( y ( x t - 1 , x t - 2 , w MAP ) - d t - 1 ) ∑ t = 2 T C t , t where C t,t is a t th element of a diagonal of a C matrix, the C matrix comprising a T by T matrix configured to represent a coupling between time series elements of the training dataset.
This invention relates to a method for computing a confidence metric in time series analysis, particularly for evaluating the accuracy of predictions in sequential data models. The method addresses the challenge of quantifying the reliability of predictions in systems where data points are temporally dependent, such as financial markets, sensor networks, or speech recognition. The confidence metric, denoted as η, is derived from a mathematical formulation that assesses the consistency between predicted and observed values over time. The formula involves summing contributions from each time step, weighted by a coupling matrix C, which captures the interdependencies between time series elements. Specifically, η is calculated as 100 times the sum of terms involving C_t,t (the diagonal elements of C) and a function Φ, which measures the discrepancy between predicted and actual values. The predicted values are generated using a model parameterized by w_MAP, derived from a training dataset, and the discrepancy is compared to a reference value d_t-1 from the previous time step. The coupling matrix C is a T-by-T matrix, where T is the number of time steps, and it encodes the strength of relationships between different time points in the sequence. The confidence metric η provides a normalized score (scaled to 100) that reflects the overall reliability of the model's predictions across the entire time series. This approach enables users to assess prediction confidence in applications where temporal dependencies are critical.
14. The method of claim 13 , further including calculating a confidence score based on the confidence metric, the calculating the confidence score comprising: providing the training dataset having a time series of samples; selecting a subset of samples from the time series of samples of the training dataset; calculating the confidence metric for the selected subset of samples; calculating an average value for the confidence metric for the selected samples; and designating the calculated average value as the confidence score for the disturbance rejection model.
This invention relates to improving the reliability of disturbance rejection models used in time series analysis. The problem addressed is the lack of a quantitative measure to assess the confidence or reliability of such models when processing real-world data, which often contains noise and disturbances. The invention provides a method to calculate a confidence score for a disturbance rejection model by analyzing a training dataset containing a time series of samples. The method involves selecting a subset of samples from the time series, calculating a confidence metric for this subset, and then determining an average value of the confidence metric across the selected samples. This average value is designated as the confidence score for the disturbance rejection model. The confidence metric itself is derived from evaluating the model's performance in rejecting disturbances within the selected subset of samples. By averaging the confidence metric over multiple samples, the method provides a robust and reliable confidence score that reflects the model's overall performance in handling disturbances. This approach ensures that the confidence score is not skewed by outliers or isolated errors, making it a more accurate indicator of the model's reliability in practical applications.
15. The method of claim 14 , wherein the confidence score comprises a percentage, the percentage indicating a likelihood that the predicted sign is correct.
This invention relates to systems for predicting and validating handwritten signatures, addressing the challenge of accurately verifying the authenticity of signatures in digital or scanned documents. The method involves analyzing a handwritten signature to generate a predicted sign, which is then compared to a reference signature to determine its validity. A confidence score, expressed as a percentage, quantifies the likelihood that the predicted sign is correct, providing a measurable assessment of the prediction's reliability. The confidence score is derived from analyzing multiple features of the signature, such as stroke patterns, pressure points, and temporal data, to enhance accuracy. The system may also incorporate machine learning models trained on a dataset of verified signatures to improve prediction performance. By assigning a confidence score, the method helps users or automated systems assess the trustworthiness of the predicted signature, reducing errors in authentication processes. This approach is particularly useful in applications requiring high-security verification, such as financial transactions, legal documents, and identity verification. The invention ensures that signature predictions are not only accurate but also accompanied by a quantifiable measure of certainty, enhancing overall system reliability.
16. The method of claim 14 , wherein the confidence score comprises a spectrum of values within which: a value of 100% indicates that the disturbance rejection model predicts a correct sign of the value change for 100% of the samples within the selected samples from the training dataset; and a value of 50% indicates that the disturbance rejection model has no confidence in predicting the correct sign of the value change for the selected samples from the training dataset.
This invention relates to a disturbance rejection model used in predictive systems, particularly for evaluating the model's confidence in predicting the correct sign of value changes in a dataset. The model processes a training dataset to assess its predictive accuracy, focusing on whether it correctly identifies the direction (positive or negative) of value changes. The confidence score is a spectrum ranging from 0% to 100%, where 100% indicates perfect accuracy in predicting the correct sign for all samples in the selected training data, while 50% signifies no confidence, meaning the model performs no better than random chance. The model may use statistical or machine learning techniques to analyze the training data, comparing predicted and actual value changes to compute the confidence score. This approach helps validate the model's reliability in real-world applications where accurate sign prediction is critical, such as in financial forecasting, sensor data analysis, or control systems. The confidence score provides a quantitative measure of the model's performance, enabling users to assess its trustworthiness before deployment.
17. The method of claim 1 , further including calculating a a confidence score based on the confidence metric, the calculating the confidence score comprising: providing the training dataset having a time series of samples; calculating the confidence metric for each of the samples of the training dataset; calculating an average value for the confidence metric for the samples; and designating the average value as the confidence score for the disturbance rejection model.
This invention relates to improving the reliability of disturbance rejection models used in time-series data analysis. The problem addressed is the lack of a quantitative measure to assess the confidence or reliability of such models when processing real-world data, which often contains noise and disturbances. The invention provides a method to calculate a confidence score for a disturbance rejection model by analyzing a training dataset containing a time series of samples. For each sample in the dataset, a confidence metric is computed, which quantifies the model's ability to reject disturbances. The individual confidence metrics are then averaged across all samples to produce a single confidence score. This score serves as an objective measure of the model's performance, allowing users to evaluate its reliability in different operating conditions. The method ensures that the confidence score is derived from empirical data, providing a robust and data-driven assessment of the model's disturbance rejection capabilities. This approach enhances decision-making by offering a clear, quantifiable metric for model performance, particularly in applications where accurate disturbance rejection is critical, such as industrial process monitoring, sensor data analysis, and predictive maintenance.
18. The method of claim 14 , further comprising: recommending a modification to the disturbance rejection model based upon a comparison of the confidence score to a predetermined minimum threshold; wherein the modification relates to a need for augmenting the training dataset of the disturbance rejection model with additional data when the confidence score fails to satisfy the predetermined minimum threshold.
This invention relates to improving disturbance rejection models used in control systems, particularly when the models lack confidence in their predictions. Disturbance rejection models are trained to identify and mitigate disturbances in dynamic systems, such as industrial processes or robotic control, but may produce low-confidence outputs when encountering novel or poorly represented disturbances. The invention addresses this by monitoring the model's confidence score—a metric indicating the certainty of its disturbance predictions. When this score falls below a predefined threshold, the system recommends augmenting the model's training dataset with additional data to improve its performance. This ensures the model adapts to previously unseen disturbances, enhancing robustness and accuracy over time. The method involves continuous evaluation of the model's confidence and proactive data collection to refine its training, reducing errors in disturbance rejection. The solution is particularly useful in applications where real-time adaptation is critical, such as autonomous systems or industrial automation, where unmitigated disturbances can lead to inefficiencies or failures. By dynamically identifying gaps in the model's training data, the invention enables self-improvement without manual intervention, improving system reliability.
19. The method of claim 6 , wherein the training comprises calculating updated values for each of the weight vector and the feedback coefficient of the network by minimizing an error function that includes a first hyperparameter, which comprises a vector for penalizing the weight vector, and a second hyperparameter, which comprises a scalar; wherein the computing the updated values for each of the weight vector and the feedback coefficient includes a matrix version of the disturbance rejection model in which a C matrix comprises a T by T matrix configured to represent a coupling between time series elements of the training dataset; wherein the error function, which is denoted as E below, that is minimized to calculate the updated values of the weight vector and the feedback coefficient is defined as: E = 1 2 w T ( Diag ( α ) ) w + 1 2 β ϵ T C ϵ where: the C comprises the C matrix; the α comprises the first hyperparameter; the β comprises the second hyperparameter; the w T comprises a transposition of the weight vector (w); the Diag(α) comprises a matrix with a diagonal element equal to a value of a corresponding diagonal element of the vector of the first hyperparameter while all other values in the matrix of the Diag (α) are set to a value of 0; the ϵ comprises an error vector that comprises a difference between the predicted values of the system output by the disturbance rejection model and the corresponding actual values of the system output at the times of the training dataset; and the ϵ T comprises a transposition of the error vector.
The invention relates to a method for training a disturbance rejection model in time series forecasting or system control applications. The method addresses the challenge of accurately modeling and rejecting disturbances in dynamic systems by optimizing both the weight vector and feedback coefficient of the network. During training, the method calculates updated values for these parameters by minimizing an error function that incorporates two hyperparameters: a vector (α) for penalizing the weight vector and a scalar (β) for scaling the error term. The error function, denoted as E, combines these hyperparameters with a matrix (C) representing the coupling between time series elements. The C matrix is a T-by-T matrix where T is the time series length. The error function is defined as E = 1/2 * w^T * (Diag(α)) * w + 1/2 * β * ϵ^T * C * ϵ, where w is the weight vector, ϵ is the error vector (difference between predicted and actual system outputs), and Diag(α) is a diagonal matrix derived from the first hyperparameter. The method uses this formulation to improve the model's ability to reject disturbances while maintaining stability and accuracy in predictions. The approach is particularly useful in applications requiring robust time series forecasting or control under noisy or dynamic conditions.
20. A system comprising: a power generating unit, including at least one of a gas turbine and a steam turbine; a control system operably connected to the power generating unit for controlling an operation thereof, the control system comprising: a hardware processor; and a machine readable storage medium on which is stored: a disturbance rejection model; and instructions that cause the hardware processor to execute a process related to control of the power generating unit; wherein the disturbance rejection model models the operation of the power generating unit so to generate a predicted value for an output of the power generating unit at a future time, the disturbance rejection model comprising a network for mapping inputs of the power generating unit to the output; wherein the process comprises: training the disturbance rejection model per a training dataset; and calculating a confidence metric for the disturbance rejection model, wherein the confidence metric is configured to indicate a probability that a predicted sign of a value change between the predicted value for the output at the future time and a measured value of the output at a previous time occurring prior to the future time is correct; wherein the predicted sign includes a positive value if the predicted value for the output at the future time increases from the measured value of the output at the previous time; and the predicted sign includes a negative value if the predicted value for the output at the future time decreases from the measured value of the output at the previous time, and wherein the control system controls the operation of the power generating unit based on the calculated confidence metric.
The system relates to power generation control, specifically for gas or steam turbines, addressing the challenge of maintaining stable and efficient operation despite disturbances. The system includes a power generating unit (gas or steam turbine) and a control system with a hardware processor and machine-readable storage. The control system uses a disturbance rejection model to predict future output values of the power generating unit by mapping input variables to output values. The model is trained using a training dataset, and a confidence metric is calculated to assess the likelihood that the predicted sign of output changes (increase or decrease) is accurate. The control system adjusts the power generating unit's operation based on this confidence metric, ensuring reliable performance. The predicted sign is positive if the output is expected to increase and negative if it is expected to decrease, providing a clear indication of trends. This approach enhances disturbance rejection and operational stability in power generation systems.
21. The system of claim 20 , wherein the training dataset comprising a time series dataset of actual values for the inputs and the output as determined from measurements taken of the operation of the power generating unit during a historical operating period, the historical operating period occurring prior to the future time; and wherein the training dataset comprises T samples, the T samples taken at times=[1, 2, . . . ,t, . . . ,T] wherein a time (t), a time (t−1), a time (t−2) denote respective samples in the training dataset where the time (t−2) occurs just prior to the time (t−1) and the time (t−1) occurs just prior to the time (t).
The invention relates to a system for training predictive models using historical operational data from power generating units. The system addresses the challenge of accurately forecasting future performance by leveraging time-series data collected during past operational periods. The training dataset includes actual measured values for input variables and output variables, recorded at discrete time intervals. The dataset consists of T samples, each corresponding to a specific time point (t), where t ranges from 1 to T. The samples are sequentially ordered such that time (t−2) precedes time (t−1), which in turn precedes time (t). This temporal structure allows the system to capture dependencies between consecutive measurements, enabling the model to learn patterns and trends over time. The historical data, collected prior to the prediction period, serves as the foundation for training machine learning models that can subsequently predict future operational states or performance metrics of the power generating unit. The sequential nature of the dataset ensures that the model can account for temporal relationships, improving the accuracy of predictions for future time points.
22. The system of claim 21 , wherein the disturbance rejection model comprises a disturbance rejection configuration in which the predicted value made by the disturbance rejection model for the output at the future time is based upon a predicted value made by the network for the output at the future time and an error multiplied by a feedback coefficient; and wherein the error comprises a difference between: a predicted value made by the network of the output at the previous time; and the measured value of the output at the previous time, wherein the measured value is based upon a measurement taken by a sensor disposed in the power generating unit for measuring an operating parameter related to the output.
This invention relates to a system for improving disturbance rejection in power generating units, particularly in controlling and optimizing their performance by reducing the impact of disturbances. The system includes a disturbance rejection model that predicts future output values of the power generating unit. The model generates these predictions by combining a predicted output value from a neural network with an error term, which is adjusted by a feedback coefficient. The error term is calculated as the difference between the neural network's predicted output at a previous time and the actual measured output at that same time. The measured output is obtained from a sensor within the power generating unit, which monitors an operating parameter directly related to the output. This feedback mechanism allows the system to dynamically adjust predictions, enhancing accuracy and stability in the presence of disturbances. The approach leverages real-time sensor data to refine predictions, ensuring more reliable control and performance optimization of the power generating unit.
23. The system of claim 22 , wherein the network comprises a neural network that includes multiple layers having nodes, the multiple layers including at least an input layer, an output layer, one or more hidden layers, and forward weight matrixes; and wherein: the input layer comprises a plurality of the nodes, the plurality of the nodes corresponding respectively to the inputs, wherein each of the plurality of the nodes is configured to receive an input signal relating to a value for a particular one of the inputs; the output layer comprises at least one of the nodes, the at least one of the nodes corresponding to the output; the one or more hidden layers are disposed between the input layer and the output layer, each of the one or more hidden layers comprising a plurality of the nodes; and the forward weight matrices comprise connectors that connect the nodes of successive layers of the multiple layers of the neural network and a weight value for each of the connectors; wherein a weight vector defines the weight values for the connectors of the forward weight matrices.
This invention relates to neural network systems designed for processing input data to generate an output. The system addresses the challenge of efficiently structuring and training neural networks to handle complex data transformations. The neural network comprises multiple interconnected layers, including an input layer, an output layer, and one or more hidden layers. Each layer contains nodes that process and transmit signals through the network. The input layer nodes receive input signals corresponding to specific input values, while the output layer nodes produce the final output. Hidden layers, positioned between the input and output layers, further process the data through their nodes. The network uses forward weight matrices to connect nodes between successive layers, with each connection assigned a weight value. These weight values are collectively defined by a weight vector, which determines the strength of each connection. The system leverages these weighted connections to propagate signals through the network, enabling the neural network to learn and adapt based on input-output relationships. This structure allows for efficient data processing and pattern recognition in various applications.
24. The system of claim 23 , wherein the confidence metric is based on a probabilistic distribution calculated for the T samples of the training dataset; wherein the inputs at the time (t−1) and the time (t−2) are described by input vectors, which are denoted, respectively, as an input vector (x t−1 ) and an input vector (x t−2 ); wherein the output at the time (t−1) is denoted as a output (d t−1 ); wherein the step of calculating the probabilistic distribution for the predicted value of the output at the time (t) comprises calculating a mean (μ) and a variance (σ 2 ) that define a normal distribution (N).
This invention relates to a system for predicting outputs in a time-series dataset using probabilistic modeling. The system addresses the challenge of accurately forecasting future values in sequential data by incorporating uncertainty quantification through probabilistic distributions. The system processes a training dataset containing T samples, where each sample includes input vectors and corresponding outputs at different time steps. Specifically, the system uses input vectors from prior time steps (denoted as x(t-1) and x(t-2)) to predict an output at a future time step (t). The prediction is based on a probabilistic distribution, which is calculated by determining the mean (μ) and variance (σ²) of the predicted output values. This distribution, typically a normal distribution (N), provides a confidence metric that quantifies the uncertainty of the prediction. The system leverages this probabilistic approach to improve the reliability of time-series forecasts by accounting for variability in the data. The invention is particularly useful in applications where understanding prediction uncertainty is critical, such as financial forecasting, weather prediction, or industrial process monitoring.
27. The system of claim 20 , wherein the processor is configured to execute the process further comprising calculating a confidence score based on the confidence metric for the disturbance rejection model, the calculating the confidence score including: providing a training dataset having a time series of samples; selecting a subset of samples from the time series of the training dataset; calculating the confidence metric for each of the selected subset of samples; calculating an average value for the confidence metric for the selected samples; and designating the calculated average value as the confidence score for the disturbance rejection model.
This invention relates to a system for evaluating the performance of a disturbance rejection model used in time series data analysis. The system addresses the challenge of assessing the reliability of models that filter or reject disturbances in time series data, ensuring accurate predictions or control actions. The system includes a processor configured to calculate a confidence score for the disturbance rejection model. The process involves using a training dataset containing a time series of samples. A subset of samples is selected from this time series, and a confidence metric is calculated for each sample in the subset. The average of these confidence metrics is then computed and designated as the confidence score for the model. This score quantifies the model's ability to reject disturbances effectively, providing a measurable indicator of its performance. The system enhances decision-making by offering a standardized way to evaluate model reliability in real-world applications.
Unknown
February 18, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.